Add TypedOneHotEncoder

[manuzhang]

This adds a typed API over Spark ML's OneHotEncoderEstimator since Spark 2.3.0

[codecov-io]

Codecov Report

Merging #322 into master will increase coverage by 1.53%. The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #322      +/-   ##
==========================================
+ Coverage   94.83%   96.36%   +1.53%     
==========================================
  Files          53       56       +3     
  Lines         968      991      +23     
  Branches        9        9              
==========================================
+ Hits          918      955      +37     
+ Misses         50       36      -14

Impacted Files	Coverage Δ
...cala/frameless/ml/feature/TypedOneHotEncoder.scala	100% <100%> (ø)
...main/scala/frameless/ops/RelationalGroupsOps.scala	79.16% <0%> (-18.46%)	:arrow_down:
...la/frameless/functions/NonAggregateFunctions.scala	100% <0%> (ø)	:arrow_up:
...re/src/main/scala/frameless/CatalystAbsolute.scala
...re/src/main/scala/frameless/CatalystBitShift.scala	100% <0%> (ø)
.../frameless/CatalystNumericWithJavaBigDecimal.scala	100% <0%> (ø)
core/src/main/scala/frameless/CatalystRound.scala	100% <0%> (ø)
...aset/src/main/scala/frameless/ops/GroupByOps.scala	98.36% <0%> (+31.69%)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 012f1a1...b51b337. Read the comment docs.

[imarios]

@manuzhang thanks for the PR! Will be looking at this shortly (I hope!).

[OlivierBlanvillain]

@manuzhang I had a quick look at the diff and it looks good to me, but this is an area for frameless I'm not very familiar with.

@atamborrino Would you have time for a review?

[imarios]

@manuzhang I am taking a closer look at the PR. I think this implementation can be made better and more type-safe. A nice API would be:

val ds: TypedDataset[(String, Int, Long)] = ...
ds.transformOneHot(ds('_1)): TypedDataset[(String, Int, Long, Vector)]

The above will not compile if you try to do one-hot encoding to any column that is not String or Char, so you can add further type-safety restriction on the type of column you want to transform. Also, note how the type of the resulting dataset correctly has a new column of type vector appended to the end.

[manuzhang]

@imarios TypedOneHotEncoder is a TypedEstimator which generates a type-safe TypedTransformer with fit. The transformer from TypedOneHotEncoder requires Int as inputs and appended Vector in outputs which are checked at compile time.

[imarios]

@manuzhang I totally dropped the ball on this PR. Let me go over this one more time and merge the PR.